Real-time multi-object tracking on highly dynamic environments

An accurate representation of the environment is important to plan tasks and efficiently control an autonomous robot. Multi-object tracking has long been an important task in robotics research whenever an autonomous robot needs to plan its tasks on an environment with multiple moving objects. Specially when the application is defined by real-time constraints in a fast-paced environment, a trade-off between accuracy and performance is automatically imposed. This paper presents a solution for real-time multi-object tracking on stochastic and highly dynamic environments. Although not limited to this specific application domain, the implementation and evaluation of this solution was performed on a team of autonomous soccer-playing robots. The proposed solution is detailed from the detection and feature extraction phase to the general problem of efficiently tracking objects in the environment. Results on the real testbed are also presented and discussed.